It's your post that is nonsense, as it appears you've read about half of what I wrote before you replied.
> What was the state of the gas and brake, what was the steering angle, how long was it from detection to action. Those things concretely answer the question of "would a human have done better or worse in this situation?"
No, "those things" you just identified in "the telemetry" can not "concretely answer" the question of "would a human have done better."
In particular, that's because "better" includes detecting the obstacle or risk sooner than the computer did. (Not merely time between detection and braking.) To answer that, other humans must scrutinize the video, which means the video is important.
That'll be my last reply though, I'm watching a youtube stream of a guy trying to get gentoo to boot off NTFS, it's not going well, but it is more interesting than providing detailed analysis of simple language that you've misunderstood.
> Did I at any point say the video wasn't important?
Sounds like Motte and Bailey time.
The expansive Bailey: "We don't need raw video to concretely answer whether the machine did better or worse than a human, the telemetry is sufficient."
The defensible Motte: "I never said the video was unimportant..."
> pedantic [...] detailed analysis
Step 1: Casually dismiss the other person as having not read your post.
Step 2: When they lay it out in detail, whine that they read your post too closely.
What most people want to compare is time from when the bicyclist could have reasonably been detected to action. This requires looking at the video frames to decide when we think a human would have figured it out.
It sounds like you're making a deeply unreasonable assumption that no human could have detected or anticipated the cyclist faster than the computer did.
Given the current state of the art, those times are frequently not the same, and the difference cannot be answered without seeing the human-consumable video footage of a prospective human driver.